Pluralis Research

Pluralis Research is an AI research organization working on decentralized, model-parallel training of large foundation models. Its central concept, termed Protocol Learning, is a training framework in which a model is split across geographically separated devices connected only over the internet, so that no single participant ever holds a full copy of the model weights. The model is described as remaining inside the protocol — trainable and usable, but not extractable — which the company presents as a basis for programmatic value flow to contributors and for collective ownership of the resulting models.

The lab publishes peer-reviewed research and technical reports on the underlying methods. Published work includes asynchronous pipeline-parallel optimization using a Nesterov-style correction for stale updates (ICML 2025), Subspace Networks for communication-efficient model parallelism (NeurIPS 2025, reporting an 8B LLaMA model trained with transformer blocks split across four locations over standard internet links), Unextractable Protocol Models covering collaborative training and inference without weight materialization (NeurIPS 2025), Mixtures of Subspaces for bandwidth-efficient context-parallel training (NeurIPS 2025), Factored Gossip DiLoCo on reducing blocking communication in DiLoCo (ICML 2026), and architecture warm-up for stable Transformer training (ICLR 2026). A 2026 technical report describes Agora, a decentralized training system that shards models across heterogeneous internet-connected participants with fault tolerance, and the associated Pluralis-8B pretraining run.

The company runs open, multi-participant training runs that outside contributors can join, and maintains a public blog documenting engineering work including a fault-tolerant multi-party distributed training stack, post-training pipelines, and reinforcement-learning post-training on consumer hardware.

Founding story

Pluralis was founded in 2024 by Alexander Long, who serves as founder and CEO. Long holds a doctorate in computer science and worked at Amazon for more than three years as an AI engineer. He assembled a team of seven other computer scientists, all with doctorates or postdoctoral experience, including researchers from Amazon, who had worked together for a number of years prior. At the time of the March 2025 funding announcement, the company had no product, and investors described the raise as a bet on the team's ability to solve the research problem of training portions of a model, rather than an entire model, on any single machine.

Business model

Pluralis operates as a research lab funded by venture capital, developing protocol-level infrastructure for decentralized model training in which contributors supply compute to shared training runs and models are collectively owned rather than held by a single entity. Its funding round was raised as equity with a warrant for future cryptocurrency should the company choose to launch a token.

The company had no product at the time of its March 2025 funding announcement; the stated design intent of Protocol Learning is that value flows programmatically to participants who contribute to training, creating economic incentives to join the network.

Traction

Pluralis reports an open pretraining run, Pluralis-8B, of an 8.6B-parameter model trained on 500B tokens across 330 contributor nodes using its Agora system. It has papers accepted at ICML 2025, NeurIPS 2025 (three), ICLR 2026 and ICML 2026, and has released a library for fault-tolerant multi-party distributed training. A July 2026 blog post reports post-training an 8B mixture-of-experts model with GRPO on 14 consumer Macs and one B200, with held-out pass@1 on agentic search rising from 29% to 63%.

Latest developments

In 2026 the company published the Agora technical report on permissionless internet-scale pretraining and the Pluralis-8B run, released details of its multi-party training stack (January 2026), replicated a full post-training pipeline on a single node (March 2026), and reported reinforcement-learning post-training experiments on consumer Mac hardware (July 2026). It is conducting ongoing open multi-participant training runs.